Merge branch 'big-refactor' into wmt

lm_eval/api/samplers.py

@@ -48,7 +48,9 @@ class Sampler:
                     )
                     + self.target_delimiter
                     + (
-                        self.doc_to_target(doc)
+                        self.doc_to_target(doc)[0]
+                        if type(self.doc_to_target(doc)) is list
+                        else self.doc_to_target(doc)
                         if (
                             self.config.doc_to_choice is None
                             or type(self.doc_to_target(doc)) is str

lm_eval/api/task.py

@@ -771,7 +771,7 @@ class ConfigurableTask(Task):
             print(type(doc_to_text))
             raise TypeError
 
-    def doc_to_target(self, doc: dict) -> Union[int, str]:
+    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
 
         if self.prompt is not None:
             doc_to_target = self.prompt
@@ -790,8 +790,12 @@ class ConfigurableTask(Task):
                 target_string = utils.apply_template(doc_to_target, doc)
                 if target_string.isdigit():
                     return ast.literal_eval(target_string)
+                elif (target_string[0] == "[") and (target_string[-1] == "]"):
+                    return ast.literal_eval(target_string)
                 else:
                     return target_string
+        elif type(doc_to_target) == list:
+            return doc_to_target
         elif callable(doc_to_target):
             return doc_to_target(doc)
         # Used when applying a Promptsource template
@@ -1002,7 +1006,7 @@ class ConfigurableTask(Task):
                 choices = self.doc_to_choice(doc)
                 gold = choices[gold]
 
-            for key in self._metric_fn_list.keys():
+            for metric in self._metric_fn_list.keys():
                 result = results[0]
                 if self.multiple_target:
                     # in the case where we have multiple targets,
@@ -1011,18 +1015,18 @@ class ConfigurableTask(Task):
                     scores = []
                     for gold_option in gold:
                         try:
-                            result_score = self._metric_fn_list[key](
+                            result_score = self._metric_fn_list[metric](
                                 references=[gold_option],
                                 predictions=[result],
-                                **self._metric_fn_kwargs[key],
+                                **self._metric_fn_kwargs[metric],
                             )
                         except TypeError:  # TODO: this is hacky and I don't want to do it
-                            result_score = self._metric_fn_list[key](
+                            result_score = self._metric_fn_list[metric](
                                 [gold_option, result]
                             )
                         if isinstance(result_score, dict):
                             # TODO: this handles the case where HF evaluate returns a dict.
-                            result_score = result_score[key]
+                            result_score = result_score[metric]
                         scores.append(result_score)
                     if any(scores):
                         result_score = 1.0
@@ -1030,17 +1034,17 @@ class ConfigurableTask(Task):
                         result_score = 0.0
                 else:
                     try:
-                        result_score = self._metric_fn_list[key](
+                        result_score = self._metric_fn_list[metric](
                             references=[gold],
                             predictions=[result],
-                            **self._metric_fn_kwargs[key],
+                            **self._metric_fn_kwargs[metric],
                         )
-                    except TypeError:
-                        result_score = self._metric_fn_list[key]([gold, result])
+                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
+                        result_score = self._metric_fn_list[metric]([gold, result])
                 if isinstance(result_score, dict):
                     result_dict.update(result_score)
                 else:
-                    result_dict[key] = result_score
+                    result_dict[metric] = result_score
         else:
             raise ValueError(
                 f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",

lm_eval/filters/__init__.py

@@ -8,6 +8,7 @@ FILTER_REGISTRY = {
     "regex": extraction.RegexFilter,
     "majority_vote": selection.MajorityVoteFilter,
     "take_first_k": selection.TakeKFilter,
+    "remove_whitespace": extraction.WhitespaceFilter,
     # TODO: implement this filter. either it should take in an arbitrary "scoring"/reward function
     # that takes an input and returns a scalar and then should select the max reward,
     # or should implement different filters for different ways of handling a reward model's inference.

lm_eval/filters/extraction.py

@@ -36,3 +36,26 @@ class RegexFilter(Filter):
         # print(filtered_resps)
 
         return filtered_resps
+
+
+class WhitespaceFilter(Filter):
+    """ """
+
+    def __init__(self):
+        pass
+
+    def apply(self, resps):
+        def filter_set(inst):
+
+            filtered_resp = []
+            for resp in inst:
+                if resp.startswith(" "):
+                    resp = resp[1:]
+
+                filtered_resp.append(resp)
+
+            return filtered_resp
+
+        filtered_resps = [filter_set(resp) for resp in resps]
+
+        return filtered_resps

lm_eval/tasks/README.md

@@ -3,7 +3,7 @@ This list keeps track of which tasks' implementations have been ported to YAML /
 
 Boxes should be checked iff tasks are implemented in the refactor and tested for regression. Tasks should be struck through if checked *against original introducing paper* implementation or popularizing implementation. (WIP) Denotes that there exists a PR or person working on this task already.
 
-- [ ] Glue (Lintang)
+- [x] Glue
 - [x] SuperGlue
 - [ ] CoQA (Lintang)
 - [ ] DROP (Lintang)
@@ -18,7 +18,7 @@ Boxes should be checked iff tasks are implemented in the refactor and tested for
 - [x] SciQ
 - [ ] QASPER
 - [x] QA4MRE
-- [ ] TriviaQA (Lintang)
+- [x] TriviaQA
 - [x] AI2 ARC
 - [x] LogiQA
 - [x] HellaSwag
@@ -45,16 +45,16 @@ Boxes should be checked iff tasks are implemented in the refactor and tested for
 - [ ] Translation (WMT) suite (Hailey)
 - [x] Unscramble
 - [x] ~~Pile (perplexity)~~
-- [ ] BLiMP (Lintang)
+- [x] BLiMP
 - [x] ToxiGen
-- [ ] StoryCloze (Lintang)
+- [x] StoryCloze
 - [ ] NaturalQs (Hailey)
 - [x] CrowS-Pairs
 - [x] XCopa
 - [ ] BIG-Bench (Hailey)
-- [ ] XStoryCloze (Lintang)
+- [x] XStoryCloze
 - [x] XWinograd
-- [ ] PAWS-X (Lintang)
+- [x] PAWS-X
 - [x] XNLI
 - [ ] MGSM (Lintang)
 - [ ] SCROLLS

lm_eval/tasks/glue/README.md

+# GLUE
+
+### Paper
+
+Title: `GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding`
+
+Abstract: https://openreview.net/pdf?id=rJ4km2R5t7
+
+The General Language Understanding Evaluation (GLUE) benchmark is a collection of
+resources for training, evaluating, and analyzing natural language understanding
+systems. GLUE consists of:
+- A benchmark of nine sentence- or sentence-pair language understanding tasks built
+on established existing datasets and selected to cover a diverse range of dataset
+sizes, text genres, and degrees of difficulty, and
+- A diagnostic dataset designed to evaluate and analyze model performance with
+respect to a wide range of linguistic phenomena found in natural language.
+
+Homepage: https://gluebenchmark.com/
+
+### Citation
+
+```
+@inproceedings{wang-etal-2018-glue,
+    title = "{GLUE}: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding",
+    author = "Wang, Alex  and
+      Singh, Amanpreet  and
+      Michael, Julian  and
+      Hill, Felix  and
+      Levy, Omer  and
+      Bowman, Samuel",
+    booktitle = "Proceedings of the 2018 {EMNLP} Workshop {B}lackbox{NLP}: Analyzing and Interpreting Neural Networks for {NLP}",
+    month = nov,
+    year = "2018",
+    address = "Brussels, Belgium",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/W18-5446",
+    doi = "10.18653/v1/W18-5446",
+    pages = "353--355",
+    abstract = "Human ability to understand language is \textit{general, flexible, and robust}. In contrast, most NLU models above the word level are designed for a specific task and struggle with out-of-domain data. If we aspire to develop models with understanding beyond the detection of superficial correspondences between inputs and outputs, then it is critical to develop a unified model that can execute a range of linguistic tasks across different domains. To facilitate research in this direction, we present the General Language Understanding Evaluation (GLUE, gluebenchmark.com): a benchmark of nine diverse NLU tasks, an auxiliary dataset for probing models for understanding of specific linguistic phenomena, and an online platform for evaluating and comparing models. For some benchmark tasks, training data is plentiful, but for others it is limited or does not match the genre of the test set. GLUE thus favors models that can represent linguistic knowledge in a way that facilitates sample-efficient learning and effective knowledge-transfer across tasks. While none of the datasets in GLUE were created from scratch for the benchmark, four of them feature privately-held test data, which is used to ensure that the benchmark is used fairly. We evaluate baselines that use ELMo (Peters et al., 2018), a powerful transfer learning technique, as well as state-of-the-art sentence representation models. The best models still achieve fairly low absolute scores. Analysis with our diagnostic dataset yields similarly weak performance over all phenomena tested, with some exceptions.",
+}
+```
+
+### Subtasks
+
+List or describe tasks defined in this folder, and their names here:
+* `task_name`: `1-sentence description of what this particular task does`
+* `task_name2`: .....
+
+### Checklist
+
+For adding novel benchmarks/datasets to the library:
+* [ ] Is the task an existing benchmark in the literature?
+  * [ ] Have you referenced the original paper that introduced the task?
+  * [ ] If yes, does the original paper provide a reference implementation? If so, have you checked against the reference implementation and documented how to run such a test?
+
+
+If other tasks on this dataset are already supported:
+* [ ] Is the "Main" variant of this task clearly denoted?
+* [ ] Have you provided a short sentence in a README on what each new variant adds / evaluates?
+* [ ] Have you noted which, if any, published evaluation setups are matched by this variant?

lm_eval/tasks/glue/cola/default.yaml

+group: glue
+task: cola
+dataset_path: glue
+dataset_name: cola
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+doc_to_text: "{{sentence}}\nQuestion: Does this sentence make sense?\nAnswer:"
+doc_to_target: label
+doc_to_choice: ["no", "yes"]
+should_decontaminate: true
+doc_to_decontamination_query: sentence
+metric_list:
+  - metric: mcc

lm_eval/tasks/glue/mnli/default.yaml

+group: glue
+task: mnli
+dataset_path: glue
+dataset_name: mnli
+output_type: multiple_choice
+training_split: train
+validation_split: validation_matched
+doc_to_text: !function utils.doc_to_text
+doc_to_target: label
+doc_to_choice: ["True", "Neither", "False"]
+metric_list:
+  - metric: acc

lm_eval/tasks/glue/mnli/mismatch.yaml

+include: default.yaml
+task: mnli_mismatch
+validation_split: validation_mismatched
+test_split: test_mismatched

lm_eval/tasks/glue/mnli/utils.py

+def doc_to_text(doc):
+    return "{}\nQuestion: {} True, False or Neither?\nAnswer:".format(
+        doc["premise"],
+        doc["hypothesis"].strip()
+        + ("" if doc["hypothesis"].strip().endswith(".") else "."),
+    )

lm_eval/tasks/glue/mrpc/default.yaml

+group: glue
+task: mrpc
+dataset_path: glue
+dataset_name: mrpc
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+doc_to_text: "Sentence 1: {{sentence1}}\nSentence 2: {{sentence2}}\nQuestion: Do both sentences mean the same thing?\nAnswer:"
+doc_to_target: label
+doc_to_choice: ["no", "yes"]
+metric_list:
+  - metric: acc
+  - metric: f1

lm_eval/tasks/glue/qnli/promptsource.yaml → lm_eval/tasks/glue/qnli/default.yaml

-group:
-  - glue-promptsource
+group: glue
 task: qnli
 dataset_path: glue
 dataset_name: qnli
 output_type: multiple_choice
 training_split: train
 validation_split: validation
-use_prompt: "promptsource:have all you need"
+doc_to_text: "{{question}}\n{{sentence}}\nQuestion: Does this response answer the question?\nAnswer:"
+doc_to_target: label
+doc_to_choice: ["yes", "no"]
 metric_list:
   - metric: acc

lm_eval/tasks/glue/qqp/default.yaml

+group: glue
+task: qqp
+dataset_path: glue
+dataset_name: qqp
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+test_split: test
+doc_to_text: "\nSentence 1: {{sentence1}}\nSentence 2: {{sentence2}}\nAnswer:"
+doc_to_target: label
+doc_to_choice: ["no", "yes"]
+metric_list:
+  - metric: acc
+  - metric: f1

lm_eval/tasks/glue/rte/default.yaml

+group: glue
+task: rte
+dataset_path: glue
+dataset_name: rte
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+doc_to_text: "{{sentence1}}\nQuestion: {{sentence2}} True or False?\nAnswer:"
+doc_to_target: label
+doc_to_choice: ["True", "False"]
+metric_list:
+  - metric: acc

lm_eval/tasks/glue/sst/default.yaml

+group: glue
+task: sst
+dataset_path: glue
+dataset_name: sst
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+doc_to_text: "{{sentence}}\nQuestion: Is this sentence positive or negative?\nAnswer:"
+doc_to_target: label
+doc_to_choice: ["negative", "positive"]
+metric_list:
+  - metric: acc

lm_eval/tasks/glue/wnli/default.yaml

+group: glue
+task: wnli
+dataset_path: glue
+dataset_name: wnli
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+doc_to_text: "{{sentence1}}\nQuestion: {{sentence2}} True or False?\nAnswer:"
+doc_to_target: label
+doc_to_choice: ["False", "True"]
+metric_list:
+  - metric: acc

lm_eval/tasks/paws-x/README.md

+# PAWS-X
+
+### Paper
+
+Title: `PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification`
+Abstract: https://arxiv.org/abs/1908.11828
+
+The dataset consists of 23,659 human translated PAWS evaluation pairs and
+296,406 machine translated training pairs in 6 typologically distinct languages.
+
+Examples are adapted from  PAWS-Wiki
+
+Prompt format (same as in mGPT):
+
+"<s>" + sentence1 + ", right? " + mask + ", " + sentence2 + "</s>",
+
+where mask is the string that matches the label:
+
+Yes, No.
+
+Example:
+
+<s> The Tabaci River is a tributary of the River Leurda in Romania, right? No, The Leurda River is a tributary of the River Tabaci in Romania.</s>
+
+Language specific prompts are translated word-by-word with Google Translate
+and may differ from the ones used by mGPT and XGLM (they do not provide their prompts).
+
+Homepage: https://github.com/google-research-datasets/paws/tree/master/pawsx
+
+
+### Citation
+
+```
+@inproceedings{yang-etal-2019-paws,
+    title = "{PAWS}-{X}: A Cross-lingual Adversarial Dataset for Paraphrase Identification",
+    author = "Yang, Yinfei  and
+      Zhang, Yuan  and
+      Tar, Chris  and
+      Baldridge, Jason",
+    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
+    month = nov,
+    year = "2019",
+    address = "Hong Kong, China",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/D19-1382",
+    doi = "10.18653/v1/D19-1382",
+    pages = "3687--3692",
+}
+```
+
+### Groups and Tasks
+
+#### Groups
+
+* `pawsx`
+
+#### Tasks
+
+* `paws_de`: German
+* `paws_en`: English
+* `paws_es`: Spanish
+* `paws_fr`: French
+* `paws_ja`: Japanese
+* `paws_ko`: Korean
+* `paws_zh`: Chinese
+
+
+### Checklist
+
+For adding novel benchmarks/datasets to the library:
+* [ ] Is the task an existing benchmark in the literature?
+  * [ ] Have you referenced the original paper that introduced the task?
+  * [ ] If yes, does the original paper provide a reference implementation? If so, have you checked against the reference implementation and documented how to run such a test?
+
+
+If other tasks on this dataset are already supported:
+* [ ] Is the "Main" variant of this task clearly denoted?
+* [ ] Have you provided a short sentence in a README on what each new variant adds / evaluates?
+* [ ] Have you noted which, if any, published evaluation setups are matched by this variant?

lm_eval/tasks/paws-x/paws_de.yaml

+# Generated by utils.py
+dataset_name: de
+doc_to_choice: '{{[sentence1+", richtig? Ja, "+sentence2, sentence1+", richtig? Nein,
+  "+sentence2]}}'
+doc_to_text: ''
+include: pawsx_template_yaml
+task: paws_de

lm_eval/tasks/paws-x/paws_en.yaml

+# Generated by utils.py
+dataset_name: en
+doc_to_choice: '{{[sentence1+", right? Yes, "+sentence2, sentence1+", right? No, "+sentence2]}}'
+doc_to_text: ''
+include: pawsx_template_yaml
+task: paws_en